考虑降水时间相关性的地面观测-雷达-卫星遥感逐时降水融合方法研究

以广东省北部山区2018年汛期强降水时段4月23—28日龙舟水、5月7—11日龙舟水和9月16—17日台风“山竹”3次典型暴雨过程的逐时降水为研究对象，研究基于XGBoost算法与地统计学理论的地面观测-前两个时刻逐时降水-雷达-卫星遥感的多源逐时降水融合模型，充分考虑相邻时刻降水的时间相关性，得到空间分辨率为1 km的逐时降水融合数据。此外，分别利用XGBoost与随机森林(RF)算法进行不考虑降水时间相关性的地面观测-雷达-卫星遥感逐时降水融合对比试验，并对试验结果进行精度评价。结果表明：(1) 在3次暴雨过程中，三种融合模型的逐时降水融合结果具有类似的空间分布；(2) 与XGBoost和RF逐时降水融合结果相比，融合了降水时间相关性的逐时降水融合结果在不同暴雨过程的准确性均有明显改进，3次暴雨过程的决定系数(R2)平均提高了7.89%和23.27%；(3) XGBoost逐时降水融合模型的精度整体上优于RF逐时降水融合模型，3次暴雨过程的R2分别提高了7.1%、4.3%和31.4%。

AN XGBOOST-BASED GEOSTATISTICAL DATA FUSION METHOD FOR INTEGRATING HOURLY GAUGE-RADAR-SATELLITE PRECIPITATIONDATA BY CONSIDERING THE TEMPORAL CORRELATIONCHARACTERISTICS OF PRECIPITATION

In this study, an hourly precipitation data fusion model was proposed based on the framework of XGBoost-based machine learning algorithm and geostatistical theory. The model integrated the multi-source hourly precipitation data, including gauge, radar, Integrated Multi-satellite Retrievals for GPM (IMERG), Global Satellite Mapping of Precipitation (GSMaP), and gridded precipitation data. The gridded precipitation data recorded in the previous two hours were analyzed to summarize the temporal correlation characteristics of precipitation in the proposed hourly precipitation data fusion model. The 1 km IMERG and GSMaP precipitation data were downscaled by using the area-to-point Kriging method. The proposed hourly precipitation data fusion model was built using hourly precipitation data from 200 rain gauges and tested with three regional rainstorm events occurred during April 23 to 28, May 7 to 11, and September 16 to 17, 2018 in northern Guangdong. The results were compared with the results estimated from the XGBoost and Random Forest (RF) algorithms without considering the temporal correlation characteristics of precipitation. The results showed that: (1) For the three regional rainstorm events, the three models reported similar spatial distribution of hourly precipitation. (2) Compared with the XGBoost and RF models, the proposed model that integrated the temporal correlation of hourly precipitation data could significantly improve the accuracy of estimated precipitation data. (3) The XGBoost model showed better performance than the RF model in capturing the nonlinear relationship between gauge precipitation data and independent variables.